deep learning for computer vision...computer vision group deep learning for computer vision...

Computer Vision Group

Deep Learning for Computer Vision

Lecturer: Dr. Laura Leal-TaixeTutors: Caner Hazirbas, Philip Häusser

Vladimir Golkov, John Chiotellis and Lingni Ma

Technische Universität MünchenComputer Vision Group

June 21, 2016

Lecturer: Dr. Laura Leal-Taixe Tutors: Caner Hazirbas, Philip Häusser Vladimir Golkov, John Chiotellis and Lingni Ma 1 / 11

Computer Vision Group

What you will learn in this course

How DL is applied to computer vision problemsPractical experience with the most successful ML methods

Artificial Neural NetworksConvolutional Neural NetworksLong short-term memory (LSTM)

Benefits/drawbacks of the methods when applied to concrete,relevant problemsPractical project experiencesPresentation skills

Lecturer: Dr. Laura Leal-Taixe Tutors: Caner Hazirbas, Philip Häusser Vladimir Golkov, John Chiotellis and Lingni Ma 2 / 11

Computer Vision Group

What you will learn in this course

How DL is applied to computer vision problemsPractical experience with the most successful ML methods

Artificial Neural NetworksConvolutional Neural NetworksLong short-term memory (LSTM)

Benefits/drawbacks of the methods when applied to concrete,relevant problemsPractical project experiencesPresentation skills

Lecturer: Dr. Laura Leal-Taixe Tutors: Caner Hazirbas, Philip Häusser Vladimir Golkov, John Chiotellis and Lingni Ma 2 / 11

Computer Vision Group

What you will learn in this course

How DL is applied to computer vision problemsPractical experience with the most successful ML methods

Artificial Neural NetworksConvolutional Neural NetworksLong short-term memory (LSTM)

Benefits/drawbacks of the methods when applied to concrete,relevant problemsPractical project experiencesPresentation skills

Lecturer: Dr. Laura Leal-Taixe Tutors: Caner Hazirbas, Philip Häusser Vladimir Golkov, John Chiotellis and Lingni Ma 2 / 11

Computer Vision Group

What you will learn in this course

How DL is applied to computer vision problemsPractical experience with the most successful ML methods

Artificial Neural NetworksConvolutional Neural NetworksLong short-term memory (LSTM)

Benefits/drawbacks of the methods when applied to concrete,relevant problemsPractical project experiencesPresentation skills

Lecturer: Dr. Laura Leal-Taixe Tutors: Caner Hazirbas, Philip Häusser Vladimir Golkov, John Chiotellis and Lingni Ma 2 / 11

Computer Vision Group

What you will learn in this course

How DL is applied to computer vision problemsPractical experience with the most successful ML methods

Artificial Neural NetworksConvolutional Neural NetworksLong short-term memory (LSTM)

Benefits/drawbacks of the methods when applied to concrete,relevant problemsPractical project experiencesPresentation skills

Lecturer: Dr. Laura Leal-Taixe Tutors: Caner Hazirbas, Philip Häusser Vladimir Golkov, John Chiotellis and Lingni Ma 2 / 11

Computer Vision Group

What you will learn in this course

How DL is applied to computer vision problemsPractical experience with the most successful ML methods

Artificial Neural NetworksConvolutional Neural NetworksLong short-term memory (LSTM)

Benefits/drawbacks of the methods when applied to concrete,relevant problemsPractical project experiencesPresentation skills

Lecturer: Dr. Laura Leal-Taixe Tutors: Caner Hazirbas, Philip Häusser Vladimir Golkov, John Chiotellis and Lingni Ma 2 / 11

Computer Vision Group

Course Structure

Three-week lecturesOne topic will be discussed each week

ANNCNNLSTM

One exercise will be assigned each week, includingpractical/theoretical questions. Solutions will be discussed inthe following weekOne-month practical project

2-3 people per group, supervised by one tutoraccess to lab computers and discussions with supervisorsduring class hours

Lecturer: Dr. Laura Leal-Taixe Tutors: Caner Hazirbas, Philip Häusser Vladimir Golkov, John Chiotellis and Lingni Ma 3 / 11

Computer Vision Group

Course Structure

Three-week lecturesOne topic will be discussed each week

ANNCNNLSTM

One exercise will be assigned each week, includingpractical/theoretical questions. Solutions will be discussed inthe following weekOne-month practical project

2-3 people per group, supervised by one tutoraccess to lab computers and discussions with supervisorsduring class hours

Lecturer: Dr. Laura Leal-Taixe Tutors: Caner Hazirbas, Philip Häusser Vladimir Golkov, John Chiotellis and Lingni Ma 3 / 11

Computer Vision Group

Course Structure

Three-week lecturesOne topic will be discussed each week

ANNCNNLSTM

One exercise will be assigned each week, includingpractical/theoretical questions. Solutions will be discussed inthe following weekOne-month practical project

2-3 people per group, supervised by one tutoraccess to lab computers and discussions with supervisorsduring class hours

Lecturer: Dr. Laura Leal-Taixe Tutors: Caner Hazirbas, Philip Häusser Vladimir Golkov, John Chiotellis and Lingni Ma 3 / 11

Computer Vision Group

Format of Final Presentation

20min presentation, 5min –10min Q&ARecommended structure

Introduction, problem definitionApproachesExperimental results and discussionsConclusions

Lecturer: Dr. Laura Leal-Taixe Tutors: Caner Hazirbas, Philip Häusser Vladimir Golkov, John Chiotellis and Lingni Ma 4 / 11

Computer Vision Group

Evaluation Criteria

Successful fulfillment of all exercisesGained expertise in the topics/projectQuality of the project presentationAttendance of classes/exercises is mandatory! In case ofsickness, medical attest is required.

Lecturer: Dr. Laura Leal-Taixe Tutors: Caner Hazirbas, Philip Häusser Vladimir Golkov, John Chiotellis and Lingni Ma 5 / 11

Computer Vision Group

Cool Projects...

FlowNet

convolutional

network

Lecturer: Dr. Laura Leal-Taixe Tutors: Caner Hazirbas, Philip Häusser Vladimir Golkov, John Chiotellis and Lingni Ma 6 / 11

Computer Vision Group

Cool Projects...

CAPTCHA Recognition with Active Deep Learning

2015 x 5415 x 54 8 x 27

4 x 1450

500 x 1378 x 1

EA

qy

B4

conv1 pool1 conv pool2

10

%

208 x 27

50

#Iterations

0 1 2 3 4 5 6 7 8

Accura

cy (

%)

0

20

40

60

80

100

added correct and uncertain samples

added random correct samples

added correct and certain samples

Lecturer: Dr. Laura Leal-Taixe Tutors: Caner Hazirbas, Philip Häusser Vladimir Golkov, John Chiotellis and Lingni Ma 7 / 11

Computer Vision Group

Cool Projects...

Biomedicine

Lecturer: Dr. Laura Leal-Taixe Tutors: Caner Hazirbas, Philip Häusser Vladimir Golkov, John Chiotellis and Lingni Ma 8 / 11

Computer Vision Group

Cool Projects...

Diffusion MRI

Lecturer: Dr. Laura Leal-Taixe Tutors: Caner Hazirbas, Philip Häusser Vladimir Golkov, John Chiotellis and Lingni Ma 9 / 11

Computer Vision Group

Study Materials

Pattern recognition and machine learning, by Christopher M.Bishop

Machine learning: a probabilistic perspective, by Kevin P.Murphy

http://www.deeplearningbook.org/ by Ian Goodfellow,Yoshua Bengio and Aaron Courville

Lecturer: Dr. Laura Leal-Taixe Tutors: Caner Hazirbas, Philip Häusser Vladimir Golkov, John Chiotellis and Lingni Ma 10 / 11